Impact of branded posts on positive sentiment on Facebook: an
investigation of the effects of social media marketing
Master Thesis
MCs in Business Administration – Marketing track University of Amsterdam
Student: Stephen Lubbersen Student Number: 11417072
Supervisor: Dr. Abhishek Nayak
Statement of Originality
This document is written by Stephen Lubbersen who declares to take full
responsibility for the contents of this document. I declare that the text and the work presented in this document are original and that no sources other than those
mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.
Table of Contents
1. Introduction ... 5
2. Literature review ... 9
2.1 Sentiment Analysis and Social media marketing ... 9
2.2 User generated content ... 9
2.3 Sentiment analysis ... 11
2.4 Importance of sentiment and positivity ... 12
2.5 Marketer and user generated content ... 14
2.6 How to create more interesting content? ... 16
2.7 Consumer engagement ... 18
2.8 Social interaction on a post level ... 20
3. Method ... 24
3.1 Introduction ... 24
3.2 Data cleaning ... 24
3.3 Independent variables ... 25
3.4 Dependent variables ... 26
3.5 Mediator and moderator ... 27
3.6 Control variables ... 27
3.7 Data ... 28
3.8 Methods ... 30
4. Results ... 31
4.1 Influence of post format and value on positive sentiment ... 31
4.2 The role of post engagement and post interaction on format of post ... 33
4.4 The role of post engagement and post interaction on value of post ... 35
4.5 The influence of control variables ... 38
5. Discussion ... 40
5.1 General discussion ... 40
5.2 Conclusion, Implications and Limitation ... 42
References ... 49 Appendix ... 60 Appendix 1: Distribution of sentiment from comments ... 60 Appendix 2: Monkeylearn Dashboard ... 60 Appendix 3: Clients Socialbakers ... 62 Appendix 4: Factorial ANOVA ... 63 Appendix 5: PROCESS Model 5 with Post Format as IV ... 65 Appendix 6: PROCESS Model 5 with Post Value as IV ... 67 Appendix 7: Questionnaire format ... 69 Appendix 8: 25 most influential brands by Mavrck ... 71 Appendix 9: PROCESS Model 5 with amount of likes as OV ... 72 Appendix 10: Less brand central post ... 74
Abstract
Social media has become an important tool for brands to persuade consumers. One specific way to persuade consumers is by posting new content in the form of branded content on various social networking sites. This study examines the role of post specific content and its influence on sentiment. I introduce the concept of
sentiment analysis since it has limited coverage within the social media landscape but shows promise due to its deeper meaning. This study is motivated by two research questions: 1. “Can consumer sentiment be used as a metric to measure branded
content?” 2. “What is the effect of various post related characteristics such as
content topic and content format on consumer sentiment with the use of sentiment
analysis?” This thesis has a sample of 102 branded posts, collected from Facebook
comments below the branded content. The findings from the research illustrate how positive sentiment is realized, consistent with previous literature we find support where videos outperform photos in positive sentiment. Contrary to current literature, entertaining posts created more positive sentiment compared to informational posts. Finally we find support that the effect of both post format and post content was moderated over time with the amount of interaction a post generated. The findings support the usage of a new metric where we see a shift from number driven metrics towards deeper meaning metrics. Managers can utilize this information by altering their metrics to better understand consumers and overall to improve their social media content strategy.
Key words: Sentiment analysis, Machine learning, Social networking sites, Consumer Behaviour, User-generated content, and Marketing communications.
1. Introduction
We are currently living in a world where it would be impossible to forget about social media. Social media, as everybody knows, has seen a huge rise in usage and is becoming increasingly important for brands (Social media week, 2016). This is mainly because consumers are turning away from traditional media and are more frequently visible on social media for their search of information and to make purchase decisions (Lempert, 2006). Around 75% of the Internet users have utilized social media to read upon experiences and this number is growing steadily (Kaplan & Haenlein, 2010). As a response managers have started heavily investing in social media to create relationships and to interact with their customers. Brands have created branded fan pages on Facebook and have started actively measuring the performance of the brand on social media(McAlexander, Schouten, & Koenig, 2002).
One of those ways to measure the performance of the brand is by capturing the reaction of the consumer, consumer-based brand equity has been recognized for its accuracy to predict brand utility from the view of a customer (Kim, Gon Kim, & An, 2003). Customer orientation has been acknowledged to be the most important driver of customer value (Blocker, Flint, Myers, & Slater, 2011). A lot of research has emphasized different ways to capture the attitude of the consumer.
A specific stream of capturing attitude comes from the user-generated content stream. A big chunk of the literature of user-generated content research has based attitude on the ratings they give to a product (Chevalier & Mayzlin, 2006; Dellarocas, Zhang, & Awad, 2007; Dhar & Chang, 2009; Ghose, Ipeirotis, & Li, 2012; Godes & Mayzlin, 2004). Within the social media stream we see the same happening, the performance of a brand is being measured with metrics such as the amount of likes
and the number of followers (De Vries, Gensler, & Leeflang, 2012; Hoffman & Fodor, 2010).
Following up on this; there seems to be a shift where there is critique on using numbers to measure performance. Reviews and ratings are positively biased and do not convey a deeper meaning (Li & Hitt, 2008). The same seems to be true for social media metrics, likes are found to not influence outcome variables such as sales (John, Mochon, Emrich, & Schwartz, 2017). Liking a brand in real life does not equal liking on social media. Whereas a consumer had to tell a story in real life about why he likes the brand; on social media he only has to push a button.
The solution to solve this gap seems to come from a more qualitative approach where the view of Fournier (2016) is supported (Fournier, Quelch, & Rietveld, 2016). A brand has to step away from data management and start emphasizing a meaning mind-set where social listening has a strong role in understanding the consumer better. Sentiment analysis has been one option for social listening and focuses on measuring an opinion and whether this opinion is neutral, positive or negative (Newman, 2016).
Sentiment analysis is quite new but is gaining track in the literature stream, sentiment analysis can predict electoral results based on political preferences (Ceron, Curini, Iacus, & Porro, 2014) and has the power to predict future stock price
movements (Oh & Sheng, 2011). Sentiment analysis is viable and consistent since it can improve the knowledge we have about consumers (Ceron et al., 2014),
unfortunately there is little known on how sentiment analysis could influence
activities by marketers on social media. Current insights are limited but are starting to grow. Research about sentiment and different social networking sites show that platforms differ from each other in their sentiment, Twitter showed more positive
sentiment compared to Facebook for instance (Schweidel & Moe, 2014; Smith, Fischer, & Yongjian, 2012). Furthermore there seems to be a general interest to further investigate post specific features, an article by Stieglitz (2013) emphasizes that a tweet with high positive sentiment results in a significantly higher amount of re-tweets by consumers (Stieglitz & Dang-Xuan, 2013).
This seems to be supported where we see more research about branded content on social media. Articles about branded content popularity emphasize on the differences in the format and the content of what is being posted, they also investigate the
influence of post length, time, engagement and other post related variables (De Vries et al., 2012; Sabate, Berbegal-Mirabent, Cañabate, & Lebherz, 2014). Unfortunately both articles use the number of likes and comments as their metric, which is earlier on discussed as being flawed. Therefore the aim of this research is to empirically
investigate what factors drive sentiment through branded content on social media, post specific.
This thesis has great theoretical contribution since it broadens our understanding of sentiment and how it could be utilized; it could also lead towards a set of new metrics that focus more on text analysis and a deeper level of analysis. This thesis emphasizes on the gap where we see a lack of literature within the social media stream focusing on deeper meaning metrics, this leads to the following two research questions.
1. Can consumer sentiment, as a base for consumer attitude, be used as a metric
to measure whether branded content does well or not?
2. What is the effect of various post related characteristics such as content topic
There is still much to discover and to understand about the usage of sentiment analysis and the influence of branded content on social media. Although machine learning sentiment analysis is able to get towards an accuracy of 80%, it still struggles to understand issues when comments include contrasts and sarcasm (Ye, Zhang, & Law, 2009).
A better understanding of this effect holds greater practical implications.
Marketers specialised in the online and social media marketing will be able to better understand the importance of sentiment analysis and how they could improve their content to create more positive and wanted behaviour. This thesis will provide a review of the relevant literature. Consequently, the research design and methods will be highlighted where the process and variables will be highlighted. The results chapter of the thesis will focus on the key results and models. Finally, the discussion and conclusion specify the significance of the findings. This thesis mainly builds on the theory of Sentiment analysis, Machine learning, Social networking sites,
2. Literature review
2.1 Sentiment Analysis and Social media marketing
As explained in the introduction the main emphasis of this thesis will be about sentiment analysis. Sentiment analysis has a lot of synonyms such like opinion mining and subjective analysis but in order to create coherence the term sentiment analysis will be used. It is important to note that there are some slight differences between these definitions, but they still fall within the same category (Liu, 2012). Sentiment analysis can be described as a mechanistic treatment of opinions, sentiment and subjectivity from text based messages (Pang & Lee, 2008). This definition will be adopted for the rest of the thesis.
Social network derives from the sociology stream where relationship between individuals are formed (Anderson, Hakansson, & Johanson, 1994). Social network, whether online or offline, is of the essence for society and businesses in order to survive and critical when competing with others (Pitt, van der Merwe, Berthon, Salehi-Sangari, & Caruana, 2006). Due to technological innovations, computers and phones now mediate interactions between individuals. Technological innovation also makes it possible to exchange messages and information across the globe in a split second (Lea, Yu, Maguluru, & Nichols, 2006). Social networking sites and the sentiment that derives from its consumers have the ability to either support or counter the effort made by brands (Michaelidou, Siamagka, & Christodoulides, 2011).
2.2 User generated content
One specific way where sentiment analysis can help companies to understand its consumers is through investigating user-generated sentiment towards the brand. This is mainly because of the rapid growth of user generated content on the web and through social media (Ghose et al., 2012). A myriad of articles have focused on the
relationship of user-generated content on sales but only use numerical data to support their evidence (Chevalier & Mayzlin, 2006; Dellarocas et al., 2007; Godes &
Mayzlin, 2004). As user-generated content is oftentimes opinion based and textual in its nature, using numerical data is biased (Li & Hitt, 2008).
User-generated content can be described as content that is created outside professional routines (Smith et al., 2012), it is related to the topic of sentiment on branded content because consumers are able to interact with each other in the comments section of branded content. Social media is a very common place where marketer- and user-generated content come together (Goh, Heng, & Lin, 2013).
Duan (2013) has investigated the user reviews for hotels using a sentiment
analysis to understand the overall evaluation. By using five different dimensions, they conclude that both negative and positive reviews are of the essence but that the
number of reviews is positive correlated with evaluation (Duan, Cao, Yu, & Levy, 2013). One of the first sources of user-generated content including social networking sites comes from Dhar (2007) who has looked into the effect of reviews on online sales for music. They come to the conclusion that a high volume of chatter to be significantly correlated with higher sales (Dhar & Chang, 2009).
Besides reviews, it has also been proven that engagement into social media brand communities leads towards a positive increase in expenditures and sales (Goh et al., 2013). In general, what all articles found is that user generated content, although methods being different, can positively be a predictor of sales, market demand and evaluation. User-generated content has the ability to influence the purchasing behaviour, marketers must therefore find ways to understand and influence what consumers are thinking and talking about (Goh et al., 2013). Sentiment analysis has
been one of those methods within the stream of user generated content to uncover sentiment, emotions and subjectivity within the text (Li & Hitt, 2008; Liu, 2012).
2.3 Sentiment analysis
Sentiment analysis is often based on two different approaches. The first is based upon a lexical approach, the lexical approach pre-tags words into a category of sentiment (Annett & Kondrak, 2008). The second approach is the machine learning based techniques; machine-learning techniques is based on supervised classification and relies on both a training data set and a test set to check for reliability (Vohra & Teraiya, 2013).
Although lexical approach has certain advantages, machine-learning techniques are known for consistently outperforming its counterpart, machine-learning
techniques are able to create higher accuracy levels (Annett & Kondrak, 2008; Vohra & Teraiya, 2013). The problem with lexical approach is the fact that there is an upper bound due to its dictionary restrictions and it is not known yet how to resolve those issues (Annett & Kondrak, 2008).
More specifically there are a couple of machine based learning techniques that stand out. Both the naïve Bayes and support vector machine techniques are often mentioned. Support vector machines are quite new and the idea is based upon risk minimization where it tries to universally learn from the data, support vector machines have been highly praised because of a high dimensional input space
(Joachims, 1998). The Naïve Bayes classifier assumes that all attributes of the data set are independent, although not often true in text classification it seems to perform quite well (McCallum & Nigam, 1998) .
While both techniques have there own advantages there seems to be a support for the usage for the support vector machine, this is mainly because support vector machines are able to create more accurate predictions, especially with big data sets (Vohra & Teraiya, 2013). This view is supported by Joachims (1998) who argues that SVM’s strength lies in a large margin approach in comparison to Naïve Bayes
(Joachims, 1998).
Within SVM and naïve Bayes there are a couple of selection features. There is a common distinction between unigrams and n-grams. Unigrams take into account single words from a document whereas n-grams emphasize on two or more words in a sequential manner (Annett & Kondrak, 2008). A paper published by Pang (2002) focuses on sentiment classification with the usage of machine learning techniques, the highest cross validation accuracy were reached when SVM was combined with either unigrams solely or a combination of unigrams and bigrams (Pang, Lee, &
Vaithyanathan, 2002).
2.4 Importance of sentiment and positivity
Through using sentiment analysis, a message will receive a certain rate of
sentiment. The stream of social media or social networking sites in combination with sentiment analysis shows the potential of what could be achieved. Sentiment analysis can predict electoral results based on political preferences (Ceron et al., 2014), it can assume revenues based on chatter (Asur & Huberman, 2010), and most importantly it shows that emotionally charged messages matter (Stieglitz & Dang-Xuan, 2013). The sentiment a brand receives on social media is of the essence.
Sentiment and word-of-mouth literature are closely related because sentiment can be seen as the outcome of word-of-mouth and electronic word-of-mouth (Jansen, Zhang, Sobel, & Chowdury, 2009b). Both negative and positive sentiment have been highlighted and appraised for their importance and therefore deserves attention since researchers do not seem to agree which one is more important.
One group of researchers emphasize that negative word-of-mouth has a more significant impact than positive sentiment(Assael, 2004; Skowronski & Carlston, 1989). Negative word-of-mouth leads towards a lower chance of repurchasing a good (Richins, 1983). This is supported within the e-commerce stream. An interesting article by Standifird (2001) focuses on the reputation of the brand and the final bid on eBay auctions. Positive ratings only had a mild influence on the closing price, only an aggregate of ten or more positive ratings did significantly influence the closing price. Negative ratings have a significant detrimental effect on the closing price. A single negative comment was found to be costly (Standifird, 2001). Negative sentiment might have a higher influence but leads towards unwanted behaviour.
There is also a group of researchers who believe that positive word-of-mouth has a higher influence, positive word-of-mouth leads to a higher amount of brand
purchases compared to negative word-of-mouth. Negative word-of-mouth is found to be less diagnostic compared to positive word-of-mouth when we are talking about the choice of a brand (East, Hammond, & Lomax, 2008).
Positive word-of-mouth is considered by the industry to be a powerful marketing tool to influence consumers (Jansen, Zhang, Sobel, & Chowdury, 2009a). Much in line, Naylor (2000) determines that positive word-of-mouth is more used in
engagements between consumers than negative word-of-mouth. This is found to be controversial since most articles at the time mainly find negative word-of-mouth to be
more influential. Furthermore it is observed that less negative word-of-mouth leads towards an increasing in the number of people they talk to (Naylor & Kleiser, 2000).
The antecedents of positive word-of-mouth are investigated by Lien (2014) who uncovers that when users trust the brand together with a right set of attitudes that are based upon entertainment, sociality and information leads towards more positive word-of-mouth (Lien & Cao, 2014). These antecedents have been backed up by Ranaweera (2003), trust and satisfaction are found to have a significant positive relationship with positive word-of-mouth (Ranaweera & Prabhu, 2003).
Positive word-of-mouth is considered to be more closely associated to comments about the quality of a service or product whereas negative word-of-mouth is linked to signs of dissatisfaction. Positive word-of-mouth is more difficult to achieve and therefore more valuable, increasing positive word-of-mouth also diminishes the amount of negative word-of-mouth (Sweeney, Soutar, & Mazzarol, 2005). Creating positivity on social media is therefore of the essence and is emphasized by most literature but negative word-of-mouth should not be forgotten due to its detrimental influence.
2.5 Marketer and user generated content
Marketer or brand-generated content is also of the essence since it conveys a message by the brand where consumers are able to express their opinions on.
Consumers are interested to engage in online brand-related content for three reasons factors; first of all a consumer wants to gather information by the content; it is also possible they want to be entertained (Muntinga, Moorman, & Smit, 2011). A third explanation could be that consumers want to consume branded content because of economical gains (Hennig-Thurau, Gwinner, Walsh, & Gremler, 2004).
When advertisement is more adapted towards the target audience, it will result into a higher liking by the consumers (Hornikx & O’Keefe, 2009). Likeable
advertisements have the power to produce more positive judgements about the brand and therefore create more positive user generated content as a result (Smit, Van Meurs, & Neijens, 2006). To add to this idea, an interesting research which focused on the differences between the two sources noted that both marketer-generated and user-generated sites are complementary to each other in terms of decision making (Bronner & De Hoog, 2010). They argue that both types of sites are being visited to form a judgement. The rise of social media has made it possible for a two-way
interaction, where brands are able to create brand related content and at the same time consumers are able to participate and express their opinion (Lim, Chung, & Weaver, 2012).
Branded social content can be beneficial but could also have detrimental effects. The quantity of social media posts have been related to a long term loss in revenue and a loss of followers (Wang, Greenwood, & Pavlou, 2017). Just liking a brand does not lead towards more meaningful or change in behaviour, rather more personal and deeper meaning to posts should be emphasized (John et al., 2017). The article by Fournier (2016) supports this movement where brands underleverage the deeper meaning of insights. A better understanding of customers should lead towards better content (Fournier et al., 2016).
For marketers it is of the essence to create interesting content, branded social content creates a new additional touch point for marketers to engage and interact with the consumer. The immediate feedback in the comments helps the brand to uncover new insights as well (Murdough, 2009). Content on branded social networking sites
therefore has to be fresh and constantly flowing, it has to engage with the audience to result into consumer liking (Ling et al., 2005).
2.6 How to create more interesting content?
Branded content is crucial and can be subdivided into a myriad of different topics, but content type has mostly been segmented into three different categories by
previous studies. Those are informational posts, which focuses on bringing product relevant information; entertainment posts where humour is the main message; and promotional posts where the main attention goes towards specific deals; and lastly social posts which want to trigger interaction by asking questions to an audience (Muntinga et al., 2011) .
As concluded by other authors, content type strongly influences the level and kind of online engagement. Entertainment posts were most likely to create a high amount of arousal, a change in attitude and return on the website (Raney, Arpan, Pashupati, & Brill, 2003).
From a mobile marketing perspective there is a similar pattern where
consumers use their phones for various reasons when consuming promotional content. Consumers have a more positive attitude towards mobile marketing and the
advertisement when it holds an entertainment value or high information value. The effect of the entertainment value is higher than the information seeking value (Bauer, Reichardt, Barnes, & Neumann, 2005). The fact that entertaining content has a higher value than informational content is also supported by Chi (2011) who explores the influence of user motivation to engage on Facebook. Both on advertisements by brands and virtual brand communities the entertaining need is higher than the informational need. It especially relates to bonding and the need for online social capital (Chi, 2011).
On the other side there is also a stream of researchers that find both entertainment and informational content to have the same value. Ducoffe (1996) argues that advertising value was depending on the perceived levels of entertainment and information. They both hold value towards consumer’s evaluation towards the ad (Ducoffe, 1995). In a recent analysis of social media advertising value, the same result can be seen where the information and entertainment are both just as important (Dao, Le, Cheng, & Chen, 2014). The following is hypothesized due to the strong arguments made in favour of entertaining content:
H1: Consumer sentiment is more positive when online marketer-generated post holds an entertainment purpose compared to a post that holds an informational purpose
Another way in which branded content can be distinguished is based on the level of vividness and interaction. Vividness can be described as the richness of a the presented information based on its formal features (Steuer, 1992). Vividness therefore depends on the characteristics of the medium; those characteristics are breadth and depth. Breadth explains the amount of sensory dimensions being presented at the same time whilst depth explains the resolution of each channel.
Low vividness is created for images, medium vividness for links towards articles and high vividness for videos (Cvijikj & Michahelles, 2013). It is being argued that a higher degree of interaction of vividness and interaction leads towards enhancing attitudes (Coyle & Thorson, 2001) but also an increase in click-through rates (Lohtia, Donthu, & Hershberger, 2003).
The effect of vividness on the number of likes is also supported de Vries (2012) where higher level of vividness leads towards a higher amount of likes (De
Vries et al., 2012). Much in line with this view, another group of researchers have investigated the factors that influence the popularity of a brand of Facebook (Sabate et al., 2014). Using some kind of visualization significantly influenced the number of comments and likes a post received.
Fortin (2005) has also done research about the role of interactivity and vividness on social presence and involvement and comes to the same kind of conclusion. Increasing the level of vividness leads to an increase of social presence, involvement and arousal. This was not the case for interactivity where only moderate interactivity leads to a higher impact; their suggestion therefore is to focus on
vividness (Fortin & Dholakia, 2005) . The following is hypothesized based on the amount of evidence where higher vividness leads towards more favourable outcomes.
H2A: Consumer sentiment will be more positive on content containing higher levels of vividness. A post containing a video creates more positive sentiment than a photo. H2B: Consumer sentiment will be more positive on content containing higher levels of vividness. A post containing a link creates more positive sentiment than a photo.
2.7 Consumer engagement
Consumer engagement is of the essence in this research since it shows a clear development in the theory for the relationship between a brand and the consumer. Consumer engagement can be described as the psychological state that occurs when consumers go through an interactive and co-creative process with a brand or product (Brodie, Hollebeek, Jurić, & Ilić, 2011). This interactive effect takes place on
different levels. Low-level consuming engagement can be described as liking a branded page or by watching and viewing a post. Low-level engagement did not create a meaningful and active behaviour but is the most common activity among
individuals (Tsai & Men, 2013). Therefore consumer engagement on social media can be described as consumers engaging in conversations on companies’ Facebook pages by commenting and asking questions, this is high-level engagement.
Consumer engagement is of the essence for brands because it creates value and leads to an increase in brand equity and loyalty (Cova, Pace, & Park, 2007). A compelling research by Dessart (2015) found that an individual is facing a double focus of participation. Besides interaction with the brand, it also interacts with the community surrounding the brand.
Customer engagement has been linked towards social benefits and economical benefits (Gummerus, Liljander, Weman, & Pihlström, 2012). Customer engagement has the potential to make fans out of ordinary customers, in order to become a fan; they have to go through stages. Those stages are connection, interaction, satisfaction, loyalty and engagement (Sashi, 2012).
On the other side, using social media just as a channel for engaging with customers could fail. Not always has customer engagement resulted in positive results and this is mainly because of a gap between what a customer wants and how the company responds (Heller Baird & Parasnis, 2011). In general, customer engagement, whether that is with the community or the brand itself creates an increase in brand loyalty (Dessart, Veloutsou, & Morgan-Thomas, 2015). Other researchers support this result as well, where customer engagement has a positive effect on perceived
relationship benefits (Gummerus et al., 2012). More specifically customer
engagement results into an increase of empowerment, connection, satisfaction and trust (Brodie, Ilic, Juric, & Hollebeek, 2013). A brand can sustain its loyalty through positive online interactions in general, specifically by giving positive customer help after a negative experience.
Equally important is to look at the antecedents for consumers to engage with brands. The perceived benefits for consumers to engage in the comments are driven by different factors. Gummerus (2012) elaborates that the benefits have a social, economical and entertainment element in it (Gummerus et al., 2012). A research by Tsai (2013) has tried to find which of these motivations drives engagement on social media the most. Remunerations was mentioned the most, followed by informational and entertainment based drivers (Tsai & Men, 2013). Also Brodie (2013) comes to the conclusion that the consumers’ need is the main reason why the engagement process is being triggered. It is furthermore being mentioned that the consumer engagement process is very interactive and experiential process, meaning that other drivers such as sharing, learning and socializing are also of the essence (Brodie et al., 2013). Since engagement depends on the kind of motivation as mentioned before, it makes it very likely that a mediation process is going on where the relationship between the topic of the post and consumer sentiment can be explained by the consumer engagement (Field, 2013).
H3: The relationship between the topic of post and overall sentiment is mediated by the customer engagement rate such that this relationship is stronger for informational posts when the customer engagement rate increases compared to the effect with entertainment posts.
2.8 Social interaction on a post level
Another important characteristic for branded content is the amount of interaction. Interactions differs from engagement since social interaction happens when
consumers interact with each other to gain feedback, share ideas, share opinions and to create new opportunities (Fischer & Reuber, 2011). Social interaction focuses on
the consumer-to-consumer relationship whereas engagement focuses on the
consumer-to-brand relationship. Social interaction therefore relates very much to the user-generated and word-of-mouth concept because user-generated content is also based on the relationship between consumers.
Social interaction is found to be moderating the relationship between social media and the way people experience news where users can choose what kind of media they are interested in and want to discuss about. Social media interaction makes it possible for the user to have discussions about what he/she finds to be important (Hermida, Fletcher, Korell, & Logan, 2012).
An article by Labrecque (2014) focuses specifically on the role of parasocial interaction, which can be described as the combination of frequency and the conversation of the persona or message it tries to send (Rubin, Perse, & Powell, 1985). They find that parasocial interaction leads towards positive outcomes such as a sense of connectedness, loyalty and willingness to provide information (Labrecque, 2014).
That social interaction leads towards positive outcomes might be due to the ability for individuals to tell about their own experience. Individuals find disclosing information about the self towards others to be intrinsically rewarding (Tamir & Mitchell, 2012), thus more social interaction should lead towards more positive sentiment.
These interactions are furthermore characterized by five primary motivations. These are entertainment, connection with the brand, service responses, incentives and promotions. Most thought-provoking is the fact that consumers interact with each other more because of the entertainment value and less for the informational value (Rohm, D. Kaltcheva, & R. Milne, 2013). This should lead to a combination of the
topic of the post and the post interaction to influence the amount of positive sentiment and therefore affirm the initial thought of the moderating role of social interaction.
H4: The positive relationship between topic of the post and positive sentiment is moderated by consumer interaction; this relationship is stronger for higher values of consumer interaction.
Also interesting is the fact that social interaction has an influence on the richness of the message. Mazzarol et al., (2007) has identified that the richness of the message plays an important role for word-of-mouth (Mazzarol, Sweeney, & Soutar, 2007). When a message was richer it resulted into more word-of-mouth. An increase of word-of-mouth has been strongly related with an increase on sales (Chevalier & Mayzlin, 2006; Duan, Gu, & Whinston, 2008; Forman, Ghose, & Wiesenfeld, 2008). The combination of an increase of richness in the post and social interaction should therefore lead to more positive behaviour in the form of sentiment, which again confirms the initial though of moderation.
H5: The positive relationship between vividness of the post and positive sentiment is moderated by consumer interaction, so that this relationship is stronger for higher values of consumer interaction.
Marketer generated content Topic of post (informational, entertainment) Consumer Sentiment Marketer generated content Format of post (link, image, video) Post engagement rate (number of engagements by brand) Post Interaction H1 H3 H4 H5 H2A H2B
All hypotheses together result in the following conceptual model:
3. Method
3.1 Introduction
The goal of this thesis is to understand the relationship between online marketer-generated content and consumer sentiment. In order to collect data, a certain amount of firms needs to be selected to gather branded content variables such as consumer engagement rate and interaction rates. A list of the 25 most influential brands on social media in the United States of America has been aided as a guideline, the list can be found in appendix 8 (Mavrck, 2016). This ensured that there would be enough comments below the posts and that the majority of the comments were in English language, popular brands also have a wider coverage of consumers, therefore having a more accurate portrait of the main population. This approach leads towards a non-probability sampling technique where brands and the posts were purposefully selected instead of randomized. Some additional brands were added based on their reputation and above-mentioned conditions. Consequently other variables and data were
included to create a master file with all the information. Comments for each post were gathered when all necessary information was available, a tool provided by
Nextanalytics made it possible to scrape the comments from Facebook (Next Analytics, 2017).
3.2 Data cleaning
The aim of this thesis is to get a general sense of the sentiment of consumers on marketer generated content. Therefore the population consists out of the active users on Facebook that are able to speak the English language, this comes to a total of 350 million people (Statista, 2017). The advice given by Barlett (2001) will be followed in terms of margin of error. Both the independent and dependent variables are
level for this thesis will be 95% and a standard deviation of 0.5 (Qualtrics, 2013). Since it is encouraged to find as much data as possible, an initial total of 56791 comments were scraped, these come from approximately 42993 different users.
Due to the amount of irrelevant comments for each post, it still had to be cleaned to create a valid data set. Therefore comments consisting only out of a tag were deleted from the dataset, a tag can be seen as a comment where only a name appears. Another set of comments was deleted due to either being unreadable, not
understandable or duplicated by the same user. A final set of comments was deleted that mainly consist out of comments made by the brand due to their forced positivity. Around 67% of the total amount of comments was used in the final data set, resulting in 38068 usable comments.
3.3 Independent variables
The first independent variable focuses on the categorization of informational versus entertainment value of a post. Two independent coders have coded the post based on a list made my Luarn (2015) who made a clear distinction between different kinds of posts (Luarn, Lin, & Chiu, 2015). Cohen’s Kappa was used to determine the amount of agreement between the two independent coders (McHugh, 2012). The informational and entertainment value show a sufficient amount of agreement (Kappa score = 0.753 & .904). The independent coders did not find agreement on five specific posts out of the total dataset. For these five posts, a third coder was introduced and aided in the discussion whether the post holds a certain value or not. The
questionnaire can be found at appendix 7.
The other dependent variable is vividness. The two main levels of vividness have been adopted from former literature (Coyle & Thorson, 2001; De Vries et al., 2012; Luarn et al., 2015). Medium levels of vividness for posts, which can be described as
events, have been removed from this research due to the lack of these posts. Posts containing a link have replaced events for medium vividness since they contain richer formal features than a photo but less compared to videos (Steuer, 1992).
3.4 Dependent variables
Sentiment will be measured on a three-point scale where a comment can either be negative, neutral or positive. The human sentiment categorization part has adopted the validated six-item seven point Likert scale to assess in which category a comment falls into (Larsen & Ketelaar, 1991). Based on their categorization, an initial amount of 5609 comments from twenty different posts by five different brands have been categorized into one of the three sentiment categories. The 5609 comments have been uploaded to Monkeylearn1 to let the machine learn what correct sentiment is. The machine achieved an accuracy level of 83% with average precision of 80.3% and average recall of 81.7%, the dashboard can be found in appendix 2 (Monkelyearn, 2017). All posts went through Monkeylearn for classification of sentiment and were all checked by humans for the needed correction. Comments that were categorized receive a probability of correctness, comments that received an 80% probability or higher were not checked, comments below 80% probability were manually checked by humans to ensure valid data. A distribution of the comments in sentiment can be found in appendix 1. Finally, all individual comments were attached to a higher level of analysis. Post level analysis is based on the aggregate number and percentage of negative, neutral and positive comments for each post.
1 Monkeylearn is a company that specializes in analyzing text through machine
learning, the classifier algorithm being used is a support vector machine (Monkelyearn, 2017).
3.5 Mediator and moderator
Post interaction and post engagement are two variables received by Social bakers2. Social bakers offers solutions for social media analytics and are known for their accurate data and variety of large brands as customers, some of their customers are highlighted in appendix 3 (Socialbakers, 2017). Post interaction is measured by the sum of interactions divided by the number of fans of the branded page.
Interactions can be seen as the total amount of likes, comments, and comments. The metric shows how much conversation the post has generated. Post engagement is a metric that measures how often the brand interacts with its customers. It is measured by the amount of times the brand has answered on comments and questions made by consumers on that specific post.
3.6 Control variables
A couple of other variables are added based on literature from different streams that might give additional explanation on the sentiment of the post. An article by de Vries (2012) about the popularity of branded content bases it on amount of likes and comments, it would be highly interesting to investigate the difference between likes, comments, shares and sentiment (De Vries et al., 2012).
Also brand level characteristics might be interesting. A list made by Overby (2006) has been used to categorize the company as either utility, hedonic motivated. The rise of the new online media has created a new stream of research for the hedonic and utilitarian motives of consumers and brands (Bridges & Florsheim, 2008;
Childers, Carr, Peck, & Carson, 2002; Lai, Kuan, Hui, & Liu, 2009; López & Ruiz, 2011)
Pöyry et al., (2013) has focused on the usage of a company Facebook page and the community participation following up on that. They found out that utilitarian
motivations lead towards browsing behaviour whereas hedonic motivations lead to higher participation and being more involved (Pöyry, Parvinen, & Malmivaara, 2013). The brand measurement was measured with the help of two independent coders, for both the utility and hedonic motivation there was a sufficient amount of agreement between the two independent coders (Kappa score = 0.738 & .751).
A couple of other variables are added due tot heir possible effect on sentiment. Brand centrality has been added due to its presence on Facebook where the brand plays a central role. It would be interesting whether this helps in terms of sentiment. Brand centrality is measured with the help of Smith (2012) who measured it as either having a central or a peripheral role (Smith et al., 2012).
An article by John et al., (2017) mentions that “boosting” your post only leads to significant meaningful behaviour. Therefore it is interesting whether this is also the case for positive sentiment (John et al., 2017). Boosting means that a post has received advertising money to increase the reach and duration of the post being visible to a higher number of consumers. The final two variables are post specific, the first emphasizes on the moment when someone posts to have a significantly influence on the amount of likes (Cvijikj & Michahelles, 2013), furthermore the length of a post has been found to have a significant influence on click through rates (Baltas, 2003). Its possible influence on positive sentiment should therefore not be underestimated.
3.7 Data
This thesis empirically investigated data of 23 different international brands that are actively posting new content on their Facebook page from the 25th of May 2017
in order to be chosen. The brands come from eight different product categories: electronics, money services, transportation & vehicles, beauty & cosmetics, drinks & beverages, wholesale, food & restaurants, sports & recreation. A total of 102 posts by brands on Facebook were gathered based on twenty-three variables. The 102 posts lead to a total of 38067 comments.
The average number of fans from all the brands together was 17,420,686
(SD=17381280.1). The lowest brand reached just over a million whereas the largest reaches over a hundred million followers. The average number of brand posts being investigated from each brand is 4.857 and each post receives on average 34.4% positive sentiment (SD=14.876). Most compelling is the fact that in general there are more positive than negative comments. It is in contrast with what Fournier (2011) has suggested where brands are not welcomed in the online community and that instead of brand building we see that firms are getting exposed for their weaknesses (Fournier & Avery, 2011). It is also in contrast with what John (2017) tries to explain in his article where we see that more posts lead towards a loss in followers and long term revenue, more positive than negative sentiment should result into more positive results (John et al., 2017).
Furthermore the data showed that we see a preference by brands to choose for either a photo or video format, both have over 40% whereas links are being utilized less than 15% of the time. Posts mostly have an entertainment value but are closely followed by informational posts. Due to the amount of posts containing more than just informational or entertainment value, a third category was added where posts can serve a dual purpose of both informational and entertaining.
Table 1: Descriptive Statistics Variables M (SD) F Relative F SD 1. Post Format 2.029 102 .923 Photo 42 41.2% Link 15 14.7% Video 45 44.1% 2. Post Value 1.863 102 .745 Informational 36 35.3% Entertainment 44 43.1% Both 22 21.1% 3. Post Interaction* 1.430 102 4.251 4. Post Engagement 6.671 102 7.882 5. Brand Essence 1.519 102 .726 Utillity 62 61.8% Hedonic 25 24.5% Both 14 13.7% 6. Sentiment Negative comments (%) 25.977 9889 13.652 Neutral comments (%) 39.612 15079 11.843 Positive comments (%) 34.411 13100 14.876 3.8 Methods
Two statistical analyses will be performed in this thesis due to the complexity in the conceptual framework. A factorial ANOVA test will be used to test the main effect of format and content on positive sentiment, factorial ANOVA is useful because it can analyse multiple independent variables and the interaction between them (Field, 2013). Consequently a regression based analysis will be performed; a tool titled PROCESS will test for mediation and moderation and checks for the influence of the independent variables (Hayes, 2017). Finally the last two models of the PROCESS analysis will include the control variables.
Variables Sum of Squares df Mean
Square F Sig. Corrected Model 6552.921 8 831.615 4.927 0.000 Intercept 78831.725 1 78831.725 467.025 0.000 Post Format 2173.803 2 1086.902 6.439 0.002 Post Value 3613.232 2 1808.116 10.712 0.000 Format * Value 1981.739 4 495.435 2.935 0.025 Error 15697.987 93 168.796 Total 141002.613 102
4. Results
4.1 Influence of post format and value on positive sentiment
For the main effect a two-way repeated-measures ANOVA will be used. The results for the first two hypotheses will be summarized in table 2. The formula for the ANOVA is based on the following where the first beta is based on the format of the post, the second beta is based on the value and the third is based on the interaction between both variables.
𝑷𝒐𝒔𝒕 𝒑𝒐𝒔 𝒔𝒆𝒏𝒕 = 𝜷𝟎 + 𝜷𝟏𝑭𝒐𝒓𝒎𝒂𝒕 + 𝜷𝟐𝑽𝒂𝒍𝒖𝒆 + 𝜷𝟑𝑰𝒏𝒕𝒆𝒓𝒂𝒄𝒕𝒊𝒐𝒏 + 𝜺)
Table 2: Outcome Factorial ANOVA DV: Percentage of positive sentiment
Notes. * R Squared = .298 (Adjusted R Squared = .237)
There was a significant main effect of the vividness (format) of the post on the percentage of positive sentiment of that post, F(2, 93) = 6.439, p< .002, 𝝎2 =.078. Bonferroni post hoc test acknowledged that the percentage of positive sentiment was significantly higher when a post contained a video compared to a photo (p <.001) but was not significant when comparing a link to either a photo or video (p<.443 & p<.682). Tables can be found in appendix 4.
25 30 35 40 45 50 55 60
Informational Entertaining Both
Po si tive S en timen t Photo Link Video
There was also a significant main effect of the value of the post on the percentage of positive sentiment, F(2,93) = 10.712, p<.000, 𝝎2 =.139. Bonferroni post hoc test acknowledged that the percentage positive sentiment was significantly higher for posts containing an entertainment value compared to posts containing an
informational value (p<.001). There was a nearly significant result where an
entertaining post creates more positive sentiment compared to posts containing both values (p<.052).
Finally there is also a significant interaction effect between the value of the post and the format of the post on the percentage of positive sentiment, F(4,93) = 2.935, p<.025, 𝝎2 =.055. This indicates that different levels of the value of a post affected different levels of format. More specifically, sentiment was similar for photos, links and videos when a post was informational or held both values. However, positive sentiment was significantly different for posts containing a photo (M=30.668), link (M=61.3) and video (M=44.416) under the entertaining condition. Photos and videos score significantly lower than links (p= <.000 & <.001) in the entertainment
condition.
Consumer Sentiment Marketer generated content Format of post Post engagement rate (number of engagements by brand) Post Interaction H2A H2B H3 H5
4.2 The role of post engagement and post interaction on format of post
The first PROCESS model takes into account the mediating role of post engagement and the moderating role of post interaction on the relationship between post format and consumer sentiment. PROCESS model 5 from Hayes (2017) has been used to investigate the relationship, the full table can be found at appendix 5 (Hayes, 2017). Bootstrapping procedure was performed for both models since it has been recommended. Table 3 shows the outcome of the first model.
The PROCESS model supports our second hypothesis where both links and videos lead to a significant increase of positive sentiment. The effect of the format on engagement a1 = 1.829 and 18.5, meaning that when the format goes from photo to link we see an increase of 1.8 on engagement. Both effects are not statistically different from the constant group, t=.083 & 1.253 and p=.930 &.213.
The effect of b1 = .054 indicates that when engagement rate goes up by one, we see an increase of .054 of positive sentiment. This is statistically different from zero, t=3.035, p=.003.
Engagement Rate (M) Positive Sentiment (Y)
Variables Coeff. SE P Coeff. SE P
Format Link 1.829 20.679 .930 9.325 3.913 .019 Format Video 18.500 14.750 .213 7.459 2.709 .007 Engagement Rate ---- ---- ---- .054 .018 .003 Constant 25.238 10.668 .019 28.734 2.021 .000 R2: .017 R2: .369 F(2,99): .867, p<.423 F(6,95): 9.269, p<.000
Effect Boot SE Boot LLCI Boot ULCI
Indirect effect 1 .099 .594 -1.111 1.368
Indirect effect 2 1.004 1.006 -.276 3.451
The indirect effect for both links and video is non significant since the lower bound is negative whereas the upper bound is positive. Therefore there is no mediation going on, instead there is a direct effect of engagement on to sentiment where higher levels of engagement lead towards more positive sentiment. The direct effect for link and the direct effect of videos are found to be significant, this is inconsistent with the factorial ANOVA, post containing a link apparently do create significantly more positive sentiment.
Table 3: Outcome Model 1
DV: Percentage of positive sentiment
The interaction effect for XM1 is c3=26.9 and is statistically different from zero, t(95)=2.603, p<.011. Thus, the effect of the post format links on sentiment depends on the amount of interaction there is. Moreover this model accounts for 36.9% of
variance in positive sentiment. A closer look into the conditional effects revealed that only links with high interaction (effect=13.865, SE=4.81, CI: 4.31 to 23.416) scored significantly higher compared to links with medium (effect=2.045, SE=3.932, CI: -5.756 to 9.856) and low interaction (effect=.658, SE=4.156, CI:-7.593 to 8.909). More interestingly, when probing the interaction between the format and sentiment is
25,000 30,000 35,000 40,000 45,000 50,000
Photo Link Video
Po si tive S en timen t Post Format Low Interaction Medium Interaction High Interaction Consumer Sentiment Post engagement rate (number of engagements by brand) Post Interaction H1 H3 H4
quite different among the interaction conditions. In other words, sentiment only significantly increases for links under the high interaction condition, such trend is the opposite for the video group where we see a decrease. There was no significant interaction effect found for posts containing a video (p < .097)
Figure 3: Interaction post format
4.4 The role of post engagement and post interaction on value of post The third model takes into account the mediating role of post engagement and the moderating role of post interaction on the relationship between post value and consumer sentiment. The full table can be found at appendix 6. Marketer generated content Topic of post
Engagement Rate (M) Positive Sentiment (Y)
Variables Coeff. SE P Coeff. SE P
Value Entertain 14.173 11.472 .339 8.542 2.655 .002 Value Both -4.264 15.469 .819 8.510 3.678 .023 Engagement Rate ---- ---- ---- .058 .017 .001 Constant 28.173 11.472 .016 27.332 2.011 .000 R2: .015 R2: .417 F(2,99): .744, p<.478 F(6,95): 11.328, p<.000
Effect Boot SE Boot LLCI Boot ULCI
Indirect effect 1 .868 1.037 -.484 3.406
Indirect effect 2 -.249 .500 -1.336 .724
The third PROCESS model shows that our first hypothesis is supported where both entertaining and posts containing both values hold significantly more positive sentiment compared tot informational posts, the factorial ANOVA only revealed that entertaining posts have significantly more positive sentiment. The effect of the value on engagement a1 = 14.872 and -4.264 means that when the value goes from informational to entertainment we see an increase of 14.87 on engagement. Both effects are not statistically different from the constant group, t=.961 & -.229 and p=.339 & .819. The effect of b1 = .058 indicating that when engagement rate goes up by one, we see an increase of .054 of positive sentiment. This is statistically different from zero, t=3.410, p= .001. The indirect effect for entertaining posts and posts holding both values are non-significant since the lower bound is negative whereas the upper bound is positive. There is no mediation of engagement, however consistent with the first model we see a direct effect of engagement on sentiment.
Table 4: Outcome Model 3
DV: Percentage of positive sentiment
For the moderation we can see that the first effect (XM1) is c3=12.6 and is statistically different from zero, t(95)= 3.981, p<.000. Thus concluding that the effect of the post value entertainment on sentiment depends on the amount of interaction.
Moreover this model explains 41.7% of the variance in positive sentiment. A closer look into the conditional effects revealed that only high interaction (84th percentile) leads to significantly more positive sentiment (effect=10.674, SE=2.66, CI: 5.386 to 15.863). The second moderation effect (XM2) is c3=29.5 and is also found to be statistically different from zero, t(95) = 2.88, p<.005. Ergo leading to the fact that posts with both values affect positive sentiment but relies upon the amount of interaction. A closer look into the conditional effect reveals the same pattern where only high interaction (84th percentile) leads to significantly more positive sentiment (effect=13.489, SE=4., CI: 3.99 to 22.99). More interestingly, when probing the interaction between the different values and sentiment is quite different among the interaction conditions. Informational posts don’t seem to change in sentiment depending on the interaction whereas entertaining and posts containing both values increase a lot from medium to high interaction.
Figure 4: Interaction post value
26,000 28,000 30,000 32,000 34,000 36,000 38,000 40,000 42,000 44,000
Informational Entertaining Both
Low Interaction Medium Interaction High Interaction
4.5 The influence of control variables
When adding the control variables, which are models two and four, we see a couple of interesting results. Most importantly there is a continuous support for H1 and H2 where both post format and post value have a significant influence on sentiment. Also H4 and H5 get support, interaction moderates the relationship between both independent variables on positive sentiment.
The control variables also seem to be of some kind of relevance. When post format was used as the independent variable; we see that posts being made in the evening have a significant positive relationship between with positive sentiment. Furthermore the brand essence has a significant positive relationship on positive sentiment, brands with a hedonic motivation receive more positive sentiment
compared to brands with a utility-based motivation. However both variables are only supported in one of the two models consisting all control variables. The fourth model, where post value was used as the independent variable, only saw brand centrality as a significant predictor for positive sentiment. In fact both models did see a significant effect of brand centrality. Posts where the brand was not centrally displayed scored higher sentiment compared to posts where the brand played a central role. An example that highlights a post consisting low brand centrality can be found in appendix 10.
Table 5: All models
Sentiment as OV
Model: 1 2 3 4
Vividness Post
Format Photo (Baseline) 28.734** 12.872*
Format Link 9.325* 7.976* Format Video 7.459** 7.392**
Value Value Informational (Baseline) 27.322** 17.480** Post Value Entertainment 8.542** 6.787*
Value Both 8.51* 6.700
Post Engagement .054** .062** .058** .066** Post Interaction 11.418* 10.884** 2.226 1.778 Int_1 26.900* 20.124* 12.632** 9.488** Int_2 -8.196 -8.87 29.501** 26.318* Part of the Day Afternoon 7.677 3.563 day Day Evening 10.339* 6.528
Day Night 9.852 7.559 Weekday -4.528 -4.593 Post Length .025 .020 Brand Hedonic 6.478* 5.000 Brand Both 2.813 4.007 Brand Centrality 6.729* 6.078* Post Nature 2.990 1.93
Notes. * Significant at p<.05, **Significant at p<.01
5. Discussion
5.1 General discussion
The first hypothesis (H1) receives support due to its consistency in different models where entertaining posts receive more positive sentiment compared to informational posts. This is in contrast with de Vries (2012) where they did not find significant differences in popularity between posts containing an entertaining / informational element or not (De Vries et al., 2012). A possible explanation for the positive sentiment is the fact that when going through social media we usually use our peripheral vision where we do not pay a lot of attention thus we do not want to
investigate information but rather watch something enjoyable (Petty & Cacioppo, 1986).
Along with the support H1, there is substantial support for H2A. Videos create significantly more positive sentiment than photos do. There is also partial support for H2B where links have significantly more positive sentiment compared to photos, both PROCESS models suggest the hypothesis to be supported but the factorial ANOVA did not find a significant difference between links and photos.
Engagement by the brand on a post did not mediate the relationship between both independent variables and the amount of positive sentiment. H3 is therefore not supported; Informational posts did not result into more engagement by the brand. It could be due to the perception of where the brand is still seen as an uninvited factor on social media, hence the reaction of consumers to not react on informational posts and thus no opportunity for the brand to engage (Fournier & Avery, 2011).
Interaction has been found to be a significant moderator between the value of the post and sentiment. H4 is supported since higher interaction leads to more positive sentiment, more specifically high interaction influences the relationship in such way that entertaining posts and posts holding both values saw a significant increase in positive sentiment compared to informational posts.
Finally, H5 has found to be partially supported where interaction on a post
moderates the relationship of format and sentiment is partially supported. There was a significant interaction between for links and sentiment where higher interaction leads towards significantly more positive sentiment. This interaction was not significantly different for the video condition. This might be due to the fact that videos are made for the general audience whereas links are more purposefully selected for a subset of customers to get interested in, when interaction increases a bigger group of this subset of customers will see the content of a link; this is in line with an article by Matz et al., (2017) who argues that targeting on psychological traits results into higher click through rates (Matz, Kosinski, Nave, & Stillwell, 2017).
One of the control variables deserves specific attention. The variable brand centrality was found to have a significant effect on positive sentiment. Posts where the brand was not the centre of the story received significantly higher likes compared to posts where the brand was central. A possible explanation is the shift from brand towards identity centrality where an identity is conscious and when a consumer can relate to it. This has been supported by Harmon-Kizer et al., (2013) who comes to the conclusion that identity centrality leads towards a higher chance of brand connections to support ones own identity (Harmon-Kizer, Kumar, Ortinau, & Stock, 2013).
5.2 Conclusion, Implications and Limitation
User-generated content and social media content are very rich and hold a vast amount of interesting and useful information. Current literature has compressed that into ratings and numbers but has forgotten the loss of a deeper meaning of what is actually mentioned. Metrics like ratings and likes do not convey everything because of their ease of usage and thus not objective opinion. To overcome this problem; sentiment analysis has been introduced to cover the current need for a deeper understanding of consumers.
Through literature it has become clear that content must be interesting in order to catch the attention of the consumer, one of those is the format in which the content is presented, does it matter if a consumer sees a photo instead of a video? Through investigating 102 different branded posts it has been able to better understand how sentiment works and in which way sentiment could be improved.
The results are very appealing and show for instance that videos outperform photos in terms of positive sentiment. Secondly, entertaining content creates more positive sentiment compared to informational content, we like to watch amusing content instead of getting informed on Facebook. Thirdly, interaction moderates the relationship, higher interaction among the consumer group in the comments section leads towards more positive sentiment.
The findings of this study offer new insights that hold great theoretical and practical contribution to existing literature on sentiment, branded content, and the usage of social media. To the best of my knowledge this is the first thesis that
investigates the relationship between Facebook branded content and sentiment while hypothesizing a mediator and a moderator.
The support for H1 means that entertainment posts create more positive sentiment than informational posts. There has been a lot of debate whether or not informational or entertainment value is of the essence. This thesis supports the view of Raney (2003) where we find entertainment to be more important (Raney et al., 2003), and collides with the view of Ducoffe (1995) who believes that the two are equally important (Ducoffe, 1995). It could still be true that informational value is also of the essence but the results show that entertaining content has a stronger positive impact on Facebook.
There is also support for H2A and partial support for H2B; in line with de Vries (2012) it is clearly visible videos create more positive sentiment compared to photos (De Vries et al., 2012). This thesis focuses on the role of vividness due to its proposed higher influence on content (Fortin & Dholakia, 2005). However, the role of
interactivity should not be forgotten, a better understanding of the interaction between interactivity and vividness would be beneficial.
There was no support for the mediating effect of H3, informational posts did not result into more engagement by the brand, which eventually should have resulted in more positive sentiment. This is the opposite of articles by Tsai (2013) and Brodie (2013) who emphasize that the main reason for consumers to engage is because of their need to retrieve information (Brodie et al., 2013; Tsai & Men, 2013). The direct effect of engagement on sentiment does show that there is a relationship; the fact that engagement has a direct effect is in line with the view of Gummerus, research should emphasis more on the direct relationship of engagement on sentiment (Gummerus et al., 2012).
Finally there is also support for H4 and partial support for H5 where interaction moderates the relationship between both main independent variables and positive sentiment. This has been in line with what other investigators have been reporting on; social interaction leads towards more trust, connectedness, and loyalty which resulted in more positivity (Chevalier & Mayzlin, 2006; Duan et al., 2008; Forman et al., 2008; Labrecque, 2014)
This thesis has attempted to fill the gap where we see a lack of literature on sentiment analysis specifically for social media content. Being able to answer both research questions results into 1. A new possible metric that helps marketers better understand what the consumer actually thinks and 2. A better understanding of the influence of post content and format on positive sentiment and how this relationship is moderated but not mediated. Sentiment analysis is of the essence because we have yet to find a metric on social media that is reliable and lets us understand the thoughts, beliefs and attitude of the consumer. De Vries (2012) argued that the effect of liking is the same as WOM and that therefore liking a post is a good metric to measure popularity (De Vries et al., 2012). The outcomes of this thesis show different results where the amount of likes as outcome variable leads to non-significant effects on both main independent variables, appendix 9. The results of this thesis help to understand the construct of sentiment as a metric but more research is needed to fully
comprehend its usefulness.
In addition to the theoretical contributions, this study furthermore holds greater managerial implications. Most importantly it tells managers how to create more positive sentiment. Videos and in some way links create more positive sentiment than photos. Firms can use this knowledge to adjust their Facebook content towards more vivid and rich content. Secondly the value of the post tells us that specifically